Cognitive machine learning -- an intelligent approach for dimensionality reduction of internet datasets

Abstract

High-dimensional data has always been a serious problem especially when the dataset has many irrelevant attributes. With the advancement of internet and cloud computing platforms, an exceptional rise has been recorded in the complexity of internet data attributes. Furthermore, in the domain of cyber security, modern data sets are highly disorganized and carry massive information to define a single event. Nonetheless, inspection of such dispersed high-dimensional data sets requires terrific human expertise and time. The contemporary machine learning techniques have great potential to deduce the relevant information from data sets, however, human cognition is always needed as an input to learning algorithms before training phase. Therefore, conventional models collapse in pruning redundant information from data sets due to the absence of a cognitive point of view. This thesis proposes a novel fractal based cognitive model to reduce the dimensionality of two different internet data sets. The aim of the proposed research is to automate raw data attributes selection using ANN and cognitive aspects. Furthermore, the overall computational complexity of the proposed model has been reduced by pruning redundant hidden neurons of ANN. Hence, experimental results demonstrate that fractal based cognitive model selects only 7 relevant attributes from a dataset of 155, and shortlists 17 attributes from another dataset of 49 attributes. Moreover, hidden neurons pruning mechanism eliminates 108 useless neurons from a single hidden layer of 154 neurons while maintaining the maximum classification accuracy of 99.2%.